Dear David Andersen, Thank you for your submission to the 2003 Internet Measurement Conference (IMC-03). We are pleased to inform you that your paper, titled "Best-Path vs. Multi-Path Overlay Routing" has been accepted as a full paper for presentation at the conference and publication in the proceedings. Congratulations! Papers went through a rigorous two-stage reviewing process. In the first stage, each paper was reviewed by at least one program committee member; after subsequent review by the program committee chair, a number of papers were rejected at that point. All papers passing the first stage were reviewed by at least three program committee members, and in some cases by external referees as well. Of the 109 papers submitted, 33 were accepted. Reviewer comments on your paper are at the end of this message. Instructions for preparing the camera-ready version of your paper will be available shortly at http://csr.bu.edu/imc-03/camera-ready.html. Please carefully take into account the enclosed comments by the reviewers when preparing the camera-ready version. Your camera-ready paper is due on August 22, 2003. To submit your paper, revisit the IMC-03 submission site at http://csr.bu.edu/imc-03/submit, go to step 3, and use the following login name and password to enter your files: Login: dga@lcs.mit.edu Password: E00-1345536670 You will need to submit both a file containing the paper source(s) and a final file in pdf or postscript format. Congratulations on having your paper accepted. We look forward to seeing you in Miami. Sincerely, Mark Crovella IMC-03 PC Chair *=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*= First reviewer's review: >>> Classification (A = strong accept, B = weak accept, C = weak reject, D = strong reject) <<< B >>> Reasons to Accept <<< It provides baseline data for subsequent investigations, and points the way to future research. >>> Reasons to Reject <<< The results are not extremely surprising, and there are details missing that would have provided more insight. >>> Are the flaws fixable? <<< Yes. >>> Summary of Review for Authors <<< This paper presents a study of loss and latency statistics in an overlay network. The results are generally not too surprising, and mostly in agreement with previous studies. The conclusions of the study depend a lot on the specific network topology. The fact that losses on alternate paths are correlated seems to say that for many pairs of paths in the study shared a common lossly link. Explicitly addressing the topology of the network (how many end systems had low-bandwidth access links) would have helped interpret these results. Without this, it's not clear how representative the results are. The main results of the study seem to be its impetus for future work; eg, work on quickly finding good alternate paths, and maximally-independent paths. Section 5 is a very interesting addition to the basic results in the paper, laying out research questions and some preliminary analyses. The comparisons between 2002 and 2003 data are hard to interpret, because the topology of the testbed changed so much over that time. =*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*= Second reviewer's review: >>> Classification (A = strong accept, B = weak accept, C = weak reject, D = strong reject) <<< B >>> Reasons to Accept <<< The problem of overlay routing in the Internet is interesting and timely. This paper compares two routing algorithms: mesh routing based on packet replication and reactive routing based on adaptive path selection, and measures the loss and latency using the two routing schemes. >>> Reasons to Reject <<< The fact of high conditional loss probability is well observed by previous work, (BOLOT et. al. and PAXSON et. al.). Also it is well known that mesh routing and reactive routing can improve the performance of end-to-end packet delivery from previous work. >>> Are the flaws fixable? <<< N/A >>> Summary of Review for Authors <<< This paper examines the correlation of the failures on the different paths. They study several existing routing mechanisms proposed for overlay network architectures, i.e., probe-based reactive routing and redundant meshing routing, to explore the independence of path failures in Internet. The conclusions of this paper are that (1) losses on alternate paths are often not independent. (2) mesh routing improves loss rate and latency on end-to-end hosts. The paper is well written. A few concens on the paper: -- The fact of high conditional loss probability is observed by previous work, (BOLOT et. al. and PAXSON et. al.). -- It is well known and not surprising that mesh routing and reactive routing can improve the performance of end-to-end packet delivery from previous work. -- It is not quite fair to compare mesh routing with reactive routing since mesh routing consume more bandwidth. Clearly, FEC might help but it still takes advantage of redundancy. =*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*= Third reviewer's review: >>> Classification (A = strong accept, B = weak accept, C = weak reject, D = strong reject) <<< B >>> Reasons to Accept <<< The paper is well motivated. How well packet replication techniques and reactive routing techniques for overlay networks perform on the current Internet is a very topical, important and open issue. The introduction and related work sections are written well and clearly outline the problem and prior understanding of the issues involved. >>> Reasons to Reject <<< The results are difficult to interpret. It is difficult to establish the generality of the results - how specific are the results to the set of nodes chosen? The conclusions from the results are somewhat obvious, but the actual percentage improvement values are a contribution. >>> Are the flaws fixable? <<< Unknown. The results may be skewed due to a choice of nodes resulting in an atypical topology or they may be representative of an overlay network deployment on the Internet. The lack of a standard deviation or variation analysis makes this determination difficult. >>> Summary of Review for Authors <<< The first finding is that loses on alternate paths are often not independent. However, it is not shown how independent the underlying paths are. Even if the direct path and an indirect overlay path are different at the overlay level, they can share underlying IP links and then naturally the losses will be correlated. The other findings, that the average loss rate is low except for some paths, that mesh routing reduces loss rate, and path selection improves mesh routing are expected. The contributions should be in the actual numbers - by how much does each scheme fair better than the other. However, it is difficult to interpret the reliability of the results. The RON2003 data set uses 30 hosts. The majority of results presented (path loss, latency) are all averages. Presumably averages over a number of samples of each pair of hosts. These values do not give a sense of the following: (a) over short time periods, how did the values change? (b) over different topologies, how do the values change? In addition to presenting results of the 30 host topology, results for all the 30choose15 topologies can be calculated and the variance of the results among these topologies will be informative. Section 4 needs some rewriting. Table 3 is confusing in relation to the list of 7 routing methods (perhaps table 3 is redundant). In fact, there are 8 routing methods (direct is not listed). Table 3 has a typo (second line should be "lat optimized path"). To avoid confusion, the same wording should be used for DD-10 as direct-direct (i.e. "2 redundant routing with packets 10ms apart on the same path"). More details are needed on how one-way latency summaries are calculated and how latency differences are used to average out errors in Section 4.1. =*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=--=*=